An AI Reported It Was 83% Sure Leftover Drums Were Hip-Hop
The model's call that the song was blues rested mostly on one guitar part, while the researcher says the rest still sounds kind of bluesy, though borderline, to his ear.
Computerphile · Why AI is like a (Clever Hans) Horse - ComputerphileTake the blues out of a blues recording, using the AI's own definition of blues, and the leftover, mostly drums and bass with the guitar turned down, is labeled hip-hop at what the model reports as 83% confidence.
The computer science researcher demonstrating the trick had split the recording into its frequencies, the individual tones that add up to make sound, and found a small slice that was enough to make the model say blues. That slice kept mainly his guitar playing, while the bass and much of the drums were gone.
The researcher admits he does not know enough about hip-hop to judge, but says he thinks most people would agree the leftover is not really hip-hop on its own. Whether the 83% figure is really confidence is debatable, but the model presents it as high certainty.
He says the models are "not listening to the music in any way that a human would understand." They are "just doing frequency analysis." He calls this shortcut learning.
Questions people ask
How do you take the blues out of a recording?
You split the sound into its frequencies and find a small set that alone is enough to make the model answer blues. Delete those frequencies from the original, and what remains is the leftover. In this case enough sound remained to send back to the classifier for another verdict.
What does the 83% number mean?
It is the certainty score the classifier reports with its label. Researchers debate whether such scores deserve to be called confidence, because models often report high certainty when they are wrong. Here it means the hip-hop label came back looking like a strong match to the model.
Does this happen with other models?
The researcher says he has tried the same removal test on many models and it always works, with the details differing by model. That suggests the issue is a general habit of learning frequency shortcuts. It does not mean every leftover gets called hip-hop.